Intelligent Signal Processing and Coordination For The Adaptive
Intelligent Signal Processing and Coordination For The Adaptive
com/chombosan
Pirathayini Srikantha and Deepa Kundur
Intelligent I
n today’s era of the Internet of Things (IoT),
the amalgamation of information and com-
munication technologies with actuating de-
Signal Processing
vices has reached all corners of the modern
world. In the context of critical infrastructures,
such as the power grid, this cyberphysical trans-
formation has permeated all system levels as evi-
and Coordination
dent in devices ranging from crucial operational
components (e.g., generators) and advanced sen-
sors [e.g., phasor measurement units (PMUs) and
programmable controllers], to consumer-centric
Smart Grid
to empower actuating devices to adaptively and
synergistically acquire data, conduct analyt-
ics, and respond to system and environmental
changes for better power-grid operations. In this
article, we demonstrate how a hierarchical signal
An overview of data-driven processing and actuation framework can enable
grid management the tractable all-encompassing coordination of
thousands of actuating power entities to main-
tain efficient operations while accounting for
Digital Object Identifier 10.1109/MSP.2018.2877001
Date of publication: 26 April 2019 physical infrastructure limits.
Power Flow
Synchronous Plant
Transformer TN ...
Home
Industry
DN
...
DG
EV/Storage
Local Communication
Channel Global Communication
Channel
Planning Tier
Abstraction
Granularity
Increasing
Increasing
Tiers
Every bus and line in the power grid is subject to con- Traditional grid operations
straints listed in (6)–(11). The state estimation of signals The traditional electric grid has been designed to accom-
generated by measurement devices, such as PMUs, will modate highly predictable power supply and demand. Thus
inform grid operators whether the system is operating far, grid operations have been well defined and mainly com-
within the physical grid limits. These constraints form the posed of contingency analysis, planning, and maintaining
feasible set S p to which the system variables must belong balance in demand/supply [3]. In contingency analysis, a
TNs
Bus
Power Line TNs transport power across geographically dispersed re-
gions that can span across hundreds of kilometers. All bus-
es in a TN operate at high voltages to minimize excessive
power losses that are typical in long lines. The ac power grid
employs three separate phases for increasing conductor effi-
Loop ciency and safety. As all three phases in a TN are balanced,
computations are based on single-phase constructs [3]. Thus,
the TN topology can be represented as a connected graph
where a single edge (power line) connects two nodes (buses).
Figure 5 presents the simplified graphical representation of a
TN consisting of 39 buses. A TN typically has a mesh struc-
FIGURE 5. A simplified graphical representation of the IEEE 39-bus New ture, as loops or cycles can exist in the power network, as
England system [15]. illustrated in Figure 5.
H8 H4 H1 C1
B33
H9
H5
H10 C3
C2
H11 C5
H2
B1
B6
H3
Feeder 1
H6
H12
H7 C4
H13
B2
H14 C6
B4
H15
B9
H20
H21
H22 C9
H25 H18
B11
H26 H19
H16
B32
B10
B8
H24
B14
C13
Feeder 2
H30
B12
H32
H33 H31
C14
B31
H34 C16
C15
H35 H38
B34
B13
H36
H37
B16
Feeder 3
B15
H39
10 kV
10/0.4 kV
H40
H49 H41
H50 H42
H51 H43
C18
B22
C21
H52 H44
H54 H45
Substation
H53
B18
B30
H55 C17
H56 C22 H46 C19
H47
B23
H48
C23 C20
Feeder 4
H57 B17
H58
B25
B20
B19
H59
H62
H63 C24
H60
H64
H61
B21
H65
C25
H66
C26
B24
H67
Feeder 5
H68 C28
C27 H73
H69
B26
H74
B29
B28
H70 H75
H71
B27
H72
S ij = / S k, (16) power-balance relations. Capturing this transience in all three
k ! Ti phases present in the system using ac variables is not straight-
forward due to the nonlinearities inherent in these sinusoidal
where j is the parent of node i (i.e., located closer to the feed- variables. To overcome this difficulty, the dq0 frame of refer-
er), Ti represents the subtree rooted at node i, and S k is the ence is used to convert the three-phase sinusoidal ac system
complex power injection into bus k. This is a linear relation. states into 2D linear variables based on a rotating frame of
Neglecting power losses in the DN enables a conservative es- reference [25]. The transience caused by the inductance and
timation of voltage drop as these are strictly negative values in capacitance present in the lines and various power components
the power-balance equations. Relaxed feasible set has the form is captured in the resulting linear ordinary differential equa-
tions in the microgrid
Slp = {Ax # B, Cx = D}, (17)
xo dq = Ax dq + B, (20)
where A, B, C, and D are constant parameters.
The second model applies convex relaxations that transform where, for notational simplicity, constant coefficients of the
the optimal coordination problem into semidefinite (SD) and/ linear state variables x dq are grouped into matrices A ! R m # 2n
or second-order cone (SOC) programs [22]. The SOC A $ B and vector B ! R m, and these represent the microgrid attri-
relaxation converts the power-balance equations into convex butes. At the steady state, the differential terms will be 0, and
quadratic inequality constraints the resulting equations are linear in terms of x:
where A ! R m # n, B ! R m, C ! R n, D ! R, E ! R k # n, Thus, these linear equations are not only convex but also
F ! R k, G ! R l # n, and H ! R l are parameters representing exact and incorporate all three phases. For the conversion from
the relaxed physical grid attributes. SD relaxation introduces the three-phase abc to the dq0 domain, Park’s transformation
convex matrix inequality constraints in lieu of the original is applied. This is based on a common rotating frame of refer-
power-balance equality relations to obtain ence [3]. Every controller in the microgrid must maintain the
same frequency and phase for the rotating frame. These values
Slp = ) / A i x $ B, Cx # D, Ex = F 3, (19)
k
are generated by crystal oscillators present in these controllers
i =1 and are synchronized by GPS signals communicated between
these controllers [25].
where A $ 0 denotes that A is positive SD, and where A 2 B de- The dq0 frame of reference converts power-balance equations
notes that A - B $ 0, B ! R m # m and A 1 ! R m # m f A k ! R m # m into linear relations. However, when transforming other limits,
are matrices scaled by individual components of x. such as voltage-magnitude constraints, into the dq0 frame, these
These convex constraints obtained using SOC and SD become nonconvex relations. To overcome issues pertaining to
relaxations are computationally more involved than the linear nonconvexity, linear approximations can be applied in a manner
DistFlow equations. However, one major advantage is that, similar to [18] to obtain
under certain conditions (e.g., radial structure and others) out-
lined in [23], these relaxations are exact in the DN. Thus, the Slp = {Ax dq = B, Cx dq # D}. (22)
tradeoffs between the linear DistFlow model and SD/SOC
relaxations are the tightness of the relaxation versus computa- Power consumers
tional efficiency. Power consumers are increasingly deploying storage sys-
tems, microgeneration systems (e.g., solar panels), and smart
Microgrids loads (e.g., EVs) in their premises, and these can be effectively
Microgrids contain generation sources and consumers. Like the coordinated by smart energy-management systems (EMSs)
DN, these directly supply electricity to consumers and are typi- [26]. Storage devices and DG systems can be combined to
cally deployed in remote communities. Microgrids can operate reduce the amount of electricity purchased from the main
in grid-connected or islanded mode. In grid-connected mode, grid. This promotes energy independence. When it is neces-
perturbations in demand/supply are absorbed by the main grid. sary to purchase power from the main grid or to conserve
In islanded mode, the microgrid, unlike the DN, is not con- energy, smart appliances can be coordinated accordingly to
nected to the main grid, and therefore, the inertia maintained maximize c onsumer comforts. The d efinition of user comfort
by bulk-generation sources is no longer present. The microgrid is unique to each consumer and can vary based on diurnal pat-
will then need to be self-sufficient and use fast-acting signal terns, seasons, and weather conditions. User comfort is associ-
processing and control mechanisms to stabilize the system dur- ated with the tolerance Pr rj to change at various c oordination
ing transient periods [24]. horizons [27]
As transience can seriously undermine the stable opera-
tion of the microgrid, it is necessary to account for these in the 0 # P rj # Pr rj, (23)
10 D
where L is the objective of the dual problem P ED , and x * and
*
8 o are optimal values of the problem. This method is called the
Power (kW)
xn
PG theoretic approach
–v Introducing discrete variables can render Pc an NP-hard prob-
lem [10]. When there are thousands of discrete variables in-
volved (e.g., in DNs with thousands of appliances), solving the
x problem as is becomes impractical, as NP-hard problems are
not scalable or tractable. Instead of directly solving Pc , apply-
FIGURE 8. A water-filling analogy [34]. ing a PG theoretic approach to transform this problem into a
Normalized Voltage
1.05
SG: α = 0.0024
600
1
400
0.95 Max Voltage
200 Min Voltage
Voltage Limits
0.9
0
0 10 20 30 40 0 5 10 15 20
Iterations Time (h)
Z
game consisting of thousands of players equipped with discrete the distribution Z in effect is the optimal solution for P ED
strategy sets resolves many of the previously mentioned issues, [35]. This can be proven by showing that the system dynam-
as discussed later in this section [35]. The main premise behind ics ensuing from these incremental revisions has a Lyapu-
the PG approach is that, when one player switches strategies, nov function, which is exactly L ^o, Z h . The existence of a
the effect on the overall system cost is incremental. If individual Lyapunov function guarantees convergence to Z *, which is
strategy revisions are made so that the overall potential/cost of the point that results in the gradient of the Lyapunov func-
the system decreases over time, then an equilibrium state will tion being zero (i.e., 2L /2z = 0) [35]. This is also the condi-
be reached eventually [36]. tion for first-order optimality, which implies that Z * is the
Z
All possible discrete strategies available in the system form solution of P ED . The convergence speed of this method is
the set Y = " y 1, f, y m ,, where y i represents the discrete independent of the size of the system, as aggregate measures
power level associated with strategy i. The central aggregating compose the variable set.
entity maintains another variable Z = " z 1, f, z m ,, where z i is Figure 9 compares the convergence of a system composed of
an aggregate variable that represents the fraction of agents in 1,000 nodes representing smart appliances over a single coor-
the population that are using strategy i. Thus, Z is a continu- dination period of 1 min via the PG and SG methods. Realistic
ous variable defined over a convex simplex [37]: demand models, appliance usage statistics, and penetration rates
have been employed for these simulations. At each iteration, the
m
Z = 'z ! Rm ; / z i = 1, z i $ 0 6 i ! m 1 .(28) central aggregating entity broadcasts a signal to the actuating
i =1 nodes. With the SG method, it is assumed that the actuation
(i.e., demand curtailment) is continuous. With the PG theoretic
D
This variable substitution can be applied to P ED , and this method, more realistic discrete strategies are used instead. The
Z
transformed problem is referred to as P ED with an objective PG theoretic method displays fast convergence and exhibits
L ^o, Z h . The central coordinating entity can now directly no oscillations. This is not the case with the SG method when
solve this dual problem to obtain o * and Z * . The main chal- a = 0.0024. Although the PG theoretic method is highly effec-
lenge now lies in achieving this optimal distribution Z * in tive for large-scale discrete coordination, it cannot be applied
the population. Every agent may have different local operat- to systems consisting of a few nodes, as the stochastic nature of
ing conditions and constraints that the central coordinating the random decision-making process will take effect and introduce
entity does not yet know. Thus, instead of directly actuat-
ing individual agents, the coordinator computes the gradient
Z
of P ED with respect to each strategy to obtain 2L /2z and Table 3. A summary of distribution decision-making topology.
broadcasts this gradient vector to all agents in the system.
Coordination Component System Convergence
The agents will revise their local actuation strategy at a
randomly selected time in a manner that accounts for local Aggregating signal SG O (n)
constraints and moves the aggregate state of the system in Water filling O (n)
PG O( Y )
the opposite direction of the gradient. This process will be
Local feasibility check Tree network O (d )
repeated until there is no incentive to switch (e.g., gradi- Mesh network O (n)
ent is 0). At this point, Nash equilibrium is achieved, and
y k +1 = argmin L t ^ x k +1, y, o k h,
erage can be used to deduce the state of the network (e.g.,
congested, stressed, healthy, and so on), which enables indi- x
vidual agents to adaptively respond to improve the general o
k +1
= o k + t ^ x k +1 - y k +1 h .
health of the system.
In Figure 11, this social networking method has been This process entails the exchange of three different sets
applied to determine the connection of EVs to the grid for of parameters with neighboring nodes at each updating
charging purposes in a 128-bus DN system. With this decen- iteration k. These sets of parameters are guaranteed to
tralized monitoring in place, it is clear that dangerous voltage converge if Pc is convex. This is indeed the case when
violations due to congestion can be prevented. convex relaxations are applied to the power-balance rela-
tions. Furthermore, the convergence rate is proportional
ADMM approach to the number of nodes n in the system [i.e., O (n)] . Fig-
Establishing optimal power flow via the ADMM approach en- ure 12 shows the ADMM method applied in the TN set-
tails the design of signals that incorporate information about ting using the linear ac approximation of power-balance
local infrastructure states (e.g., the apparent power-flow and constraints and the change in the residual ^i.e., x - y 2h
bus-voltage magnitude) by every actuating node. These sig- over one coordination interval of 10 min. Signals are
nals are then exchanged with the nodes of direct neighbors exchanged every 6 s. This system is composed of 2,736
and used to refine local actuation to increase efficiency while buses [15], and it is clear from Figure 12 that the decen-
maintaining locally inferred feasibility. Signals are repeatedly tralized coordination mechanism results in fast conver-
exchanged with peers until a global consensus regarding the gence to the optimal solution. Thus, the communication
nominal operation of the entire system is attained. of signals to neighboring nodes that contain current val-
The design of signals is composed of two steps. First, the ues computed for local, perspective, and dual variables
nonconvex power-balance relations are converted into a set of enables every node to adaptively respond to changes in
convex constraints via one of the linear ac approximations, the system in a decentralized manner while accounting
convex relations via SD/SOC, or steady-state dq0 transforma- for physical power-balance relations.
tion, as listed in the section “Tractable Physical Grid Model-
ing.” The next step will be to decompose these power-balance
equations for each node to infer local feasibility. However, as Residual at Each ADMM Iteration
power balance depends on power flowing from directly con- 20
nected nodes, it is not possible to directly decompose this 15
Residual
As the set of strategies available to each switching agent Summary of decentralized coordination
is finite, this sequence will converge within finite time to a Decentralized coordination enables individual agents, by ex-
changing signals with neighboring nodes, to adaptively re-
spond to changes and perturbations and, thereby, progressively
improve the global state of the system while adhering to local
Confidence Interval of Improvement Factor constraints. Table 4 summarizes the three decentralized meth-
1.05
ods and applications for adaptive and proactive actuation that
1 have been presented.
0.95
0.9
Independent coordination
In all of the coordination paradigms presented earlier, the
0.85 nodes communicate. The communication channel is subject to
Potential Game
0.8
FNSGA
latencies in the order of milliseconds [46]. In certain applica-
0.75 tions that require fast-acting control and response, this latency
33-3 33-5 33-7 69-5 will not be tolerable (e.g., primary control) [24]. In these sce-
DNs narios, only local measurements (e.g., bus-voltage magnitude,
current, frequency, and so on) can be used by actuating nodes
FIGURE 13. A comparison of the performance of DN topology reconfigura- for inferring the general state of the system, and actuation is,
tion algorithms. therefore, independent of external input. Specific actuation
mechanisms include those based on droop control, primal-dual
dynamics, and machine-learning techniques.
Table 4. A summary of decentralized algorithms and applications.
Short-Term Transmission/Microgrid
Total Mesh Network Short-Term
Forecast Setpoints (Decentralized) Setpoints Forecast Tier 1
Decreasing Coordination Horizon
Increasing Abstraction
Governor/Inverter/
Appliance Tier 3
(Independent)
Actuation Monitoring
Multiple Components of the Same Type